void HermitianFromEVD
( UpperOrLower uplo,
AbstractDistMatrix<F>& APre,
const AbstractDistMatrix<Base<F>>& wPre,
const AbstractDistMatrix<F>& ZPre )
{
DEBUG_CSE
typedef Base<F> Real;
DistMatrixWriteProxy<F,F,MC,MR> AProx( APre );
DistMatrixReadProxy<Real,Real,VR,STAR> wProx( wPre );
DistMatrixReadProxy<F,F,MC,MR> ZProx( ZPre );
auto& A = AProx.Get();
auto& w = wProx.GetLocked();
auto& Z = ZProx.GetLocked();
const Grid& g = A.Grid();
DistMatrix<F,MC, STAR> Z1_MC_STAR(g);
DistMatrix<F,VR, STAR> Z1_VR_STAR(g);
DistMatrix<F,STAR,MR > Z1Adj_STAR_MR(g);
const Int m = Z.Height();
const Int n = Z.Width();
A.Resize( m, m );
if( uplo == LOWER )
MakeTrapezoidal( UPPER, A, 1 );
else
MakeTrapezoidal( LOWER, A, -1 );
const Int bsize = Blocksize();
for( Int k=0; k<n; k+=bsize )
{
const Int nb = Min(bsize,n-k);
auto Z1 = Z( ALL, IR(k,k+nb) );
auto w1 = w( IR(k,k+nb), ALL );
Z1_MC_STAR.AlignWith( A );
Z1_MC_STAR = Z1;
Z1_VR_STAR.AlignWith( A );
Z1_VR_STAR = Z1_MC_STAR;
DiagonalScale( RIGHT, NORMAL, w1, Z1_VR_STAR );
Z1Adj_STAR_MR.AlignWith( A );
Adjoint( Z1_VR_STAR, Z1Adj_STAR_MR );
LocalTrrk( uplo, F(1), Z1_MC_STAR, Z1Adj_STAR_MR, F(1), A );
}
}
inline void
ReformHermitianMatrix
( UpperOrLower uplo,
DistMatrix<Complex<R>,MC,MR >& A,
const DistMatrix<R, VR,STAR>& w,
const DistMatrix<Complex<R>,MC,MR >& Z,
const RealFunctor& f )
{
#ifndef RELEASE
PushCallStack("hermitian_function::ReformHermitianMatrix");
#endif
const Grid& g = A.Grid();
typedef Complex<R> C;
DistMatrix<C,MC,MR> ZL(g), ZR(g),
Z0(g), Z1(g), Z2(g);
DistMatrix<R,VR,STAR> wT(g), w0(g),
wB(g), w1(g),
w2(g);
DistMatrix<C,MC, STAR> Z1_MC_STAR(g);
DistMatrix<C,VR, STAR> Z1_VR_STAR(g);
DistMatrix<C,STAR,MR > Z1Adj_STAR_MR(g);
DistMatrix<R,STAR,STAR> w1_STAR_STAR(g);
if( uplo == LOWER )
MakeTrapezoidal( LEFT, UPPER, 1, A );
else
MakeTrapezoidal( LEFT, LOWER, -1, A );
LockedPartitionRight( Z, ZL, ZR, 0 );
LockedPartitionDown
( w, wT,
wB, 0 );
while( ZL.Width() < Z.Width() )
{
LockedRepartitionRight
( ZL, /**/ ZR,
Z0, /**/ Z1, Z2 );
LockedRepartitionDown
( wT, w0,
/**/ /**/
w1,
wB, w2 );
Z1_MC_STAR.AlignWith( A );
Z1_VR_STAR.AlignWith( A );
Z1Adj_STAR_MR.AlignWith( A );
//--------------------------------------------------------------------//
Z1_MC_STAR = Z1;
Z1_VR_STAR = Z1_MC_STAR;
w1_STAR_STAR = w1;
// Scale Z1[VR,* ] with the modified eigenvalues
const int width = Z1_VR_STAR.Width();
const int localHeight = Z1_VR_STAR.LocalHeight();
for( int j=0; j<width; ++j )
{
const R omega = f(w1_STAR_STAR.GetLocalEntry(j,0));
C* buffer = Z1_VR_STAR.LocalBuffer(0,j);
for( int iLocal=0; iLocal<localHeight; ++iLocal )
buffer[iLocal] *= omega;
}
Z1Adj_STAR_MR.AdjointFrom( Z1_VR_STAR );
internal::LocalTrrk( uplo, (C)1, Z1_MC_STAR, Z1Adj_STAR_MR, (C)1, A );
//--------------------------------------------------------------------//
Z1Adj_STAR_MR.FreeAlignments();
Z1_VR_STAR.FreeAlignments();
Z1_MC_STAR.FreeAlignments();
SlideLockedPartitionDown
( wT, w0,
w1,
/**/ /**/
wB, w2 );
SlideLockedPartitionRight
( ZL, /**/ ZR,
Z0, Z1, /**/ Z2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
ReformNormalMatrix
( DistMatrix<Complex<R>,MC,MR >& A,
const DistMatrix<R, VR,STAR>& w,
const DistMatrix<Complex<R>,MC,MR >& Z,
const ComplexFunctor& f )
{
#ifndef RELEASE
PushCallStack("hermitian_function::ReformNormalMatrix");
#endif
const Grid& g = A.Grid();
typedef Complex<R> C;
DistMatrix<C,MC,MR> ZL(g), ZR(g),
Z0(g), Z1(g), Z2(g);
DistMatrix<R,VR,STAR> wT(g), w0(g),
wB(g), w1(g),
w2(g);
DistMatrix<C,MC, STAR> Z1_MC_STAR(g);
DistMatrix<C,VR, STAR> Z1_VR_STAR(g);
DistMatrix<C,STAR,MR > Z1Adj_STAR_MR(g);
DistMatrix<R,STAR,STAR> w1_STAR_STAR(g);
Zero( A );
LockedPartitionRight( Z, ZL, ZR, 0 );
LockedPartitionDown
( w, wT,
wB, 0 );
while( ZL.Width() < Z.Width() )
{
LockedRepartitionRight
( ZL, /**/ ZR,
Z0, /**/ Z1, Z2 );
LockedRepartitionDown
( wT, w0,
/**/ /**/
w1,
wB, w2 );
Z1_MC_STAR.AlignWith( A );
Z1_VR_STAR.AlignWith( A );
Z1Adj_STAR_MR.AlignWith( A );
//--------------------------------------------------------------------//
Z1_MC_STAR = Z1;
Z1_VR_STAR = Z1_MC_STAR;
w1_STAR_STAR = w1;
// Scale Z1[VR,* ] with the modified eigenvalues
const int width = Z1_VR_STAR.Width();
const int localHeight = Z1_VR_STAR.LocalHeight();
for( int j=0; j<width; ++j )
{
const C conjOmega = Conj(f(w1_STAR_STAR.GetLocalEntry(j,0)));
C* buffer = Z1_VR_STAR.LocalBuffer(0,j);
for( int iLocal=0; iLocal<localHeight; ++iLocal )
buffer[iLocal] *= conjOmega;
}
Z1Adj_STAR_MR.AdjointFrom( Z1_VR_STAR );
internal::LocalGemm
( NORMAL, NORMAL, (C)1, Z1_MC_STAR, Z1Adj_STAR_MR, (C)1, A );
//--------------------------------------------------------------------//
Z1Adj_STAR_MR.FreeAlignments();
Z1_VR_STAR.FreeAlignments();
Z1_MC_STAR.FreeAlignments();
SlideLockedPartitionDown
( wT, w0,
w1,
/**/ /**/
wB, w2 );
SlideLockedPartitionRight
( ZL, /**/ ZR,
Z0, Z1, /**/ Z2 );
}
#ifndef RELEASE
PopCallStack();
#endif
}
inline void
HermitianFromEVD
( UpperOrLower uplo,
DistMatrix<F>& A,
const DistMatrix<BASE(F),VR,STAR>& w,
const DistMatrix<F>& Z )
{
#ifndef RELEASE
CallStackEntry entry("HermitianFromEVD");
#endif
const Grid& g = A.Grid();
typedef BASE(F) R;
DistMatrix<F> ZL(g), ZR(g),
Z0(g), Z1(g), Z2(g);
DistMatrix<R,VR,STAR> wT(g), w0(g),
wB(g), w1(g),
w2(g);
DistMatrix<F,MC, STAR> Z1_MC_STAR(g);
DistMatrix<F,VR, STAR> Z1_VR_STAR(g);
DistMatrix<F,STAR,MR > Z1Adj_STAR_MR(g);
DistMatrix<R,STAR,STAR> w1_STAR_STAR(g);
A.ResizeTo( Z.Height(), Z.Height() );
if( uplo == LOWER )
MakeTrapezoidal( UPPER, A, 1 );
else
MakeTrapezoidal( LOWER, A, -1 );
LockedPartitionRight( Z, ZL, ZR, 0 );
LockedPartitionDown
( w, wT,
wB, 0 );
while( ZL.Width() < Z.Width() )
{
LockedRepartitionRight
( ZL, /**/ ZR,
Z0, /**/ Z1, Z2 );
LockedRepartitionDown
( wT, w0,
/**/ /**/
w1,
wB, w2 );
Z1_MC_STAR.AlignWith( A );
Z1_VR_STAR.AlignWith( A );
Z1Adj_STAR_MR.AlignWith( A );
//--------------------------------------------------------------------//
Z1_MC_STAR = Z1;
Z1_VR_STAR = Z1_MC_STAR;
w1_STAR_STAR = w1;
DiagonalScale( RIGHT, NORMAL, w1_STAR_STAR, Z1_VR_STAR );
Z1Adj_STAR_MR.AdjointFrom( Z1_VR_STAR );
LocalTrrk( uplo, F(1), Z1_MC_STAR, Z1Adj_STAR_MR, F(1), A );
//--------------------------------------------------------------------//
SlideLockedPartitionDown
( wT, w0,
w1,
/**/ /**/
wB, w2 );
SlideLockedPartitionRight
( ZL, /**/ ZR,
Z0, Z1, /**/ Z2 );
}
}
